[openib-general] Re: Some more RMPP Questions and Comments (mad_rmpp.c)

[Sean Hefty]

Hal Rosenstock wrote:

>>Hmm... I was originally thinking that it wouldn't matter if this field was set or 
>>not, but if RMPP is not active, the field is supposed to be set to 0.
> 
> And also visa versa, if RMPP is active, the field is supposed to have
> the ACTIVE bit on.

Currently, users need to set the active bit, and are responsible for 
setting most other MAD and RMPP header fields.

>>I can think of a couple of options here.  We can extend this API to 
>>indicate if RMPP is active or not.
> 
> That's what I thought at first.

The only "issue" that I see with doing this is that I tried to avoid 
having create_send_mad format the MAD.  I made an exception for the 
payload length, since the buffer size is needed to make RMPP work. 
Maybe the right solution here is to return the payload length rather 
than setting it in the MAD.

>>Have the user clear this field if RMPP is not active.
> 
> This seems fine to me and consistent with how remote_qkey is being
> handled. This should be documented. 
> 
>>  Or let the RMPP code clear the field when sending 
>>the MAD...
> 
> How would it know to do this ?

The code would know based on whether the active bit was set.  If not, 
it can clear the payload field.  I'm not as fond of this option, since 
the user is responsible for clearing/setting the other header fields.

>>It might even make sense to have calls to initialize the MAD and RMPP 
>>header information given a list of parameters.
> 
> 
> Yes, that would save each consumer from doing this explictly.

It can better ensure that all of the necessary fields are set.

> 
>>>Also from rmpp.c::ib_create_send_mad:
>>>
>>>send_buf->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
>>>
>>>Should the remote_qkey be a passed in parameter to this routine ?
>>
>>The user could override this once the MAD has been returned, but I 
>>don't object to adding this...
> 
> 
> True. Should this be added to the comments ?

Sure ... I can do that this afternoon.

- Sean